Generative AI's Impact: Reduced Study Time, Compromised Learning in Mathematics

      The rapid proliferation of generative AI tools, epitomized by chatbots like ChatGPT, has introduced a paradigm shift in how students approach academic challenges. While widely assumed to influence learning behaviors, concrete, large-scale evidence detailing these shifts and their long-term consequences has remained elusive. Recent academic research sheds light on this critical issue, revealing a significant decline in the time students dedicate to math problems and a troubling erosion of durable knowledge, particularly when learning occurs without stringent supervision.

The Unseen Shift in Student Behavior

      Prior to this research, much of the understanding regarding AI's impact on student learning came from self-report surveys, which often indicated minimal changes in academic dishonesty despite widespread intuition suggesting otherwise. These surveys are prone to biases, such as social desirability and ambiguity around what constitutes AI misuse. Complementing these are small-scale behavioral studies that hint at problematic AI usage but lack the scope and duration to quantify long-term learning outcomes. The absence of robust, quantitative evidence has left educators and administrators navigating AI policy in the dark.

      This gap is now being addressed through extensive behavioral trace data from ALEKS, a prominent adaptive mathematics learning and assessment platform utilized by millions of students annually. The study analyzed over 3.2 million learning interactions and 12.2 million assessment response-time observations over a ten-year period, effectively spanning the pre- and post-ChatGPT eras. Unlike self-reports, these interaction logs provide an objective record of student behavior, including time-on-task and problem-solving patterns. This foundational behavioral data is crucial for understanding genuine engagement with academic material.

Methodology: A Quasi-Experimental Approach

      To precisely identify the impact of generative AI, the researchers employed a clever quasi-experimental design. They exploited the inherent variation in how easily different types of math problems could be outsourced to AI. "AI-susceptible" topics primarily consisted of text-based word problems, such as those involving proportional reasoning, algebra, or rate and mixture calculations. These problems are easily transcribed into natural language prompts, making them ideal for AI assistance.

      In contrast, "AI-resistant" topics involved graph-based or plot-based problems that demanded visual interpretation and interactive manipulation within the ALEKS platform. Their complex visual and interactive nature makes them significantly harder to solve using a simple AI chatbot prompt. This differential susceptibility creates a robust within-student, within-quarter comparison, allowing the study to isolate the effects of AI from other confounding factors like platform-wide trends, cohort differences, or general calendar-time shocks.

Key Behavioral Findings: Reduced Study Time

      The analysis revealed a substantial and consistent pattern: after the release of ChatGPT, college students showed a 2.8% decline per quarter in learning time spent on AI-susceptible problems. This accumulated to a staggering 26.9% reduction over eleven quarters. High school students exhibited an even more pronounced drop of 31.3%, while middle schoolers showed a 9.0% decrease. Interestingly, Grade 5 students demonstrated no detectable change, an age gradient consistent with the likelihood of autonomous AI access and proficiency.

      This consistent reduction in time-on-task specifically for problems easily solved by AI suggests a widespread behavioral shift. Students are completing these problems faster, but the critical question remains: are they truly learning? This phenomenon has significant implications for educational platforms and institutions looking to maintain engagement and learning efficacy. Providers of advanced monitoring solutions, such as ARSA Technology, can offer AI Video Analytics to observe and analyze engagement patterns in controlled learning or assessment environments, helping educators detect unusual shifts in study behavior that might indicate inappropriate AI use.

The Crucial Role of Proctoring and Retention

      To discern whether the reduced study time reflected genuine efficiency gains or reliance on AI assistance, the study examined student behavior under proctored conditions. The findings were stark: the divergence in time-on-task between AI-susceptible and AI-resistant problems vanished entirely when students were supervised. This observation strongly suggests that the declines in study time were indeed driven by AI use, as opposed to students simply becoming more efficient learners.

      Furthermore, the study investigated the impact on durable knowledge by analyzing retention performance on randomly assigned, proctored assessment items of the same problem types. The results were concerning: students showed a cumulative 25% decline in the odds of providing a correct response on these AI-susceptible items. This indicates that while AI might help students complete tasks faster, it undermines their ability to retain and apply the knowledge independently. Secure assessment platforms can benefit from robust identity verification and proctoring technologies. Solutions like the ARSA Face Recognition & Liveness API could be integrated into such systems to ensure the verified identity of the test-taker and prevent impersonation or proxy testing.

Unpacking the "Cognitive Surrender"

      Perhaps the most compelling evidence presented by the research comes from a critical falsification test. When the same retention estimator was applied to non-proctored assessments, it produced a large, opposite-signed increase in the odds of a correct response. This reversal is highly inconsistent with any alternative explanation based on curriculum changes, platform evolution, or shifts in student cohort composition. Instead, it powerfully reinforces the conclusion that AI use is the driving mechanism behind the observed behavioral and learning outcomes.

      The researchers aptly term this phenomenon "cognitive surrender" – a population-level indicator where students delegate cognitive effort to AI, leading to an immediate gain in task completion but a long-term loss in knowledge acquisition and retention. This insight is critical for understanding the true cost of unguided AI integration into learning processes. Custom AI solutions, like those provided by ARSA's custom AI services, could be developed for educational platforms to analyze learning data, identify patterns indicative of cognitive surrender, and trigger interventions to promote active learning strategies.

Implications for Education and AI Policy

      This groundbreaking study provides some of the first large-scale behavioral and outcome evidence quantifying how generative AI has fundamentally altered student study habits and the knowledge they build. The findings carry direct and profound implications for various stakeholders:

• Educational Research: Future research must account for the pervasive influence of AI on learning behaviors and outcomes.
• Assessment Governance: Educational institutions need to re-evaluate current assessment strategies, considering whether they adequately measure true understanding in an AI-assisted world. The need for proctored, secure environments becomes paramount.
• AI Policy: Clear policies are required to guide the ethical and effective integration of AI in education, balancing its potential benefits for efficiency with the risks to deep learning and knowledge retention.
• EdTech Development: Learning platforms must innovate to create AI tools that foster, rather than circumvent, active cognitive engagement.

      This research underscores that while AI offers immense potential to enhance educational experiences, its uncritical adoption can inadvertently lead to significant learning deficits. Safeguarding the integrity of learning processes requires a proactive and informed approach.

      For enterprises and institutions navigating the complexities of AI integration, particularly where data control, privacy, and operational reliability are paramount, ARSA Technology offers advanced AI and IoT solutions. Our expertise in developing production-ready systems for security, operations, and decision intelligence, serving various industries, can be leveraged to address challenges in securing educational data and ensuring robust assessment integrity.

      To explore how ARSA Technology can help your organization implement secure and effective AI strategies, please do not hesitate to contact ARSA for a free consultation.

      **Source:** Rismanchian, S., Uzun, H., Matayoshi, J., Cosyn, E., & Kurd-Misto, E. (2026). Faster Completion, Less Learning: Generative AI Reduced Study Time on Math Problems and the Knowledge They Build. Preprint. Under review. https://arxiv.org/abs/2605.21629